Apriel Voice 27B Preview has 65.51% task success in voice tests
The open model gets voice as input and gives replies and tool calls. It needs no second dialogue model.
The authors give Apriel Voice 27B Preview, an open model for voice, reasoning and tool use. It is built on a Gemma 4 26B-A4B checkpoint, a version of a model at one point in its training. It gives replies and tool calls with no second dialogue model. In thinking mode, its task success rate on the tau-Voice test is 65.51%.
How it works
Training has two steps. The first step sets up the voice interface. The second step changes the projector and the backbone together. The authors made voice dialogues by machine as training data.
The numbers
- 98.90 on the Big Bench Audio test
- 81.13 on the Speech-BFCL test
- 51.27% on the English EVA test
The source gives no score for other models.
What is not known
This is a preview. The source does not compare the model with other models. It gives no price, no license terms and no name of the hardware.
Insight: One voice model competes with a set of parts
On 21 October 2026, IBM released voice to text models. On 22 October, OpenAI opened its Decisions API to all customers. It has a time of 100 to 180 milliseconds. Apriel does voice input, reasoning and tool calls in one model. Vendors look like they race to own the voice layer. The lowest latency can set the standard. The measures are not the same.
Sources
Posted
